Dev builds 9-month side-project agent harness with working-memory tool outputs
Spirited_Neck1858 · reddit · 2026-09-20
An indie developer shared Xtream-Agentic-Harness, a terminal agent assistant built over 9 months of spare time, motivated by AI chats degrading after 20–25 turns as old tool results fill the context.
Key design choices: tool outputs live in working memory instead of permanent history; tool-calling uses XML tags rather than native function-calling so any model works (llama.cpp, NIM, Gemini); a local ML classifier detects hallucinated tool syntax and nudges retries; every turn is zip-archived for /restore, with searchable cross-session history. Open source on GitHub.
More from coding & agent
- Codex's Audio Mode Is Underhyped, Say Developers: 'Feels Like GPT-3 Moment' — athyuttamre · 2026-09-20
- Jens Honack: Jev Is Like a Smart Switch Statement, More Dynamic Than Low-Dimensional Determinism — JensHonack · 2026-09-20
- Teknium Critiques Tamara's Jev Compaction Strategy: Valid Use Cases, Wrong Approach — Teknium · 2026-09-20
- Swapping a 13s pipeline step for a 200ms call saves thousands per month — hardimanjames · 2026-09-20
- Opus 5-built Three.js WebGPU waves deliver stunning real-time shoreline in browser — majidmanzarpour · 2026-09-20
- Jev, a 'System One' model that only makes decisions, questions how many LLM calls agents really need — ThePromptIndex · 2026-09-20